event/sw: support event ports
authorBruce Richardson <bruce.richardson@intel.com>
Thu, 30 Mar 2017 19:30:36 +0000 (20:30 +0100)
committerJerin Jacob <jerin.jacob@caviumnetworks.com>
Tue, 4 Apr 2017 17:19:51 +0000 (19:19 +0200)
Add in the data-structures for the ports used by workers to send
packets to/from the scheduler. Also add in the functions to
create/destroy those ports.

Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Signed-off-by: Harry van Haaren <harry.van.haaren@intel.com>
Acked-by: Jerin Jacob <jerin.jacob@caviumnetworks.com>
drivers/event/sw/event_ring.h [new file with mode: 0644]
drivers/event/sw/sw_evdev.c
drivers/event/sw/sw_evdev.h

diff --git a/drivers/event/sw/event_ring.h b/drivers/event/sw/event_ring.h
new file mode 100644 (file)
index 0000000..cdaee95
--- /dev/null
@@ -0,0 +1,185 @@
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright(c) 2016-2017 Intel Corporation. All rights reserved.
+ *
+ *   Redistribution and use in source and binary forms, with or without
+ *   modification, are permitted provided that the following conditions
+ *   are met:
+ *
+ *     * Redistributions of source code must retain the above copyright
+ *       notice, this list of conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright
+ *       notice, this list of conditions and the following disclaimer in
+ *       the documentation and/or other materials provided with the
+ *       distribution.
+ *     * Neither the name of Intel Corporation nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+/*
+ * Generic ring structure for passing events from one core to another.
+ *
+ * Used by the software scheduler for the producer and consumer rings for
+ * each port, i.e. for passing events from worker cores to scheduler and
+ * vice-versa. Designed for single-producer, single-consumer use with two
+ * cores working on each ring.
+ */
+
+#ifndef _EVENT_RING_
+#define _EVENT_RING_
+
+#include <stdint.h>
+
+#include <rte_common.h>
+#include <rte_memory.h>
+#include <rte_malloc.h>
+
+#define QE_RING_NAMESIZE 32
+
+struct qe_ring {
+       char name[QE_RING_NAMESIZE] __rte_cache_aligned;
+       uint32_t ring_size; /* size of memory block allocated to the ring */
+       uint32_t mask;      /* mask for read/write values == ring_size -1 */
+       uint32_t size;      /* actual usable space in the ring */
+       volatile uint32_t write_idx __rte_cache_aligned;
+       volatile uint32_t read_idx __rte_cache_aligned;
+
+       struct rte_event ring[0] __rte_cache_aligned;
+};
+
+#ifndef force_inline
+#define force_inline inline __attribute__((always_inline))
+#endif
+
+static inline struct qe_ring *
+qe_ring_create(const char *name, unsigned int size, unsigned int socket_id)
+{
+       struct qe_ring *retval;
+       const uint32_t ring_size = rte_align32pow2(size + 1);
+       size_t memsize = sizeof(*retval) +
+                       (ring_size * sizeof(retval->ring[0]));
+
+       retval = rte_zmalloc_socket(NULL, memsize, 0, socket_id);
+       if (retval == NULL)
+               goto end;
+
+       snprintf(retval->name, sizeof(retval->name), "EVDEV_RG_%s", name);
+       retval->ring_size = ring_size;
+       retval->mask = ring_size - 1;
+       retval->size = size;
+end:
+       return retval;
+}
+
+static inline void
+qe_ring_destroy(struct qe_ring *r)
+{
+       rte_free(r);
+}
+
+static force_inline unsigned int
+qe_ring_count(const struct qe_ring *r)
+{
+       return r->write_idx - r->read_idx;
+}
+
+static force_inline unsigned int
+qe_ring_free_count(const struct qe_ring *r)
+{
+       return r->size - qe_ring_count(r);
+}
+
+static force_inline unsigned int
+qe_ring_enqueue_burst(struct qe_ring *r, const struct rte_event *qes,
+               unsigned int nb_qes, uint16_t *free_count)
+{
+       const uint32_t size = r->size;
+       const uint32_t mask = r->mask;
+       const uint32_t read = r->read_idx;
+       uint32_t write = r->write_idx;
+       const uint32_t space = read + size - write;
+       uint32_t i;
+
+       if (space < nb_qes)
+               nb_qes = space;
+
+       for (i = 0; i < nb_qes; i++, write++)
+               r->ring[write & mask] = qes[i];
+
+       rte_smp_wmb();
+
+       if (nb_qes != 0)
+               r->write_idx = write;
+
+       *free_count = space - nb_qes;
+
+       return nb_qes;
+}
+
+static force_inline unsigned int
+qe_ring_enqueue_burst_with_ops(struct qe_ring *r, const struct rte_event *qes,
+               unsigned int nb_qes, uint8_t *ops)
+{
+       const uint32_t size = r->size;
+       const uint32_t mask = r->mask;
+       const uint32_t read = r->read_idx;
+       uint32_t write = r->write_idx;
+       const uint32_t space = read + size - write;
+       uint32_t i;
+
+       if (space < nb_qes)
+               nb_qes = space;
+
+       for (i = 0; i < nb_qes; i++, write++) {
+               r->ring[write & mask] = qes[i];
+               r->ring[write & mask].op = ops[i];
+       }
+
+       rte_smp_wmb();
+
+       if (nb_qes != 0)
+               r->write_idx = write;
+
+       return nb_qes;
+}
+
+static force_inline unsigned int
+qe_ring_dequeue_burst(struct qe_ring *r, struct rte_event *qes,
+               unsigned int nb_qes)
+{
+       const uint32_t mask = r->mask;
+       uint32_t read = r->read_idx;
+       const uint32_t write = r->write_idx;
+       const uint32_t items = write - read;
+       uint32_t i;
+
+       if (items < nb_qes)
+               nb_qes = items;
+
+
+       for (i = 0; i < nb_qes; i++, read++)
+               qes[i] = r->ring[read & mask];
+
+       rte_smp_rmb();
+
+       if (nb_qes != 0)
+               r->read_idx += nb_qes;
+
+       return nb_qes;
+}
+
+#endif
index 574696b..af14cfd 100644 (file)
 
 #include "sw_evdev.h"
 #include "iq_ring.h"
+#include "event_ring.h"
 
 #define EVENTDEV_NAME_SW_PMD event_sw
 #define NUMA_NODE_ARG "numa_node"
 #define SCHED_QUANTA_ARG "sched_quanta"
 #define CREDIT_QUANTA_ARG "credit_quanta"
 
+static void
+sw_info_get(struct rte_eventdev *dev, struct rte_event_dev_info *info);
+
+static int
+sw_port_setup(struct rte_eventdev *dev, uint8_t port_id,
+               const struct rte_event_port_conf *conf)
+{
+       struct sw_evdev *sw = sw_pmd_priv(dev);
+       struct sw_port *p = &sw->ports[port_id];
+       char buf[QE_RING_NAMESIZE];
+       unsigned int i;
+
+       struct rte_event_dev_info info;
+       sw_info_get(dev, &info);
+
+       /* detect re-configuring and return credits to instance if needed */
+       if (p->initialized) {
+               /* taking credits from pool is done one quanta at a time, and
+                * credits may be spend (counted in p->inflights) or still
+                * available in the port (p->inflight_credits). We must return
+                * the sum to no leak credits
+                */
+               int possible_inflights = p->inflight_credits + p->inflights;
+               rte_atomic32_sub(&sw->inflights, possible_inflights);
+       }
+
+       *p = (struct sw_port){0}; /* zero entire structure */
+       p->id = port_id;
+       p->sw = sw;
+
+       snprintf(buf, sizeof(buf), "sw%d_%s", dev->data->dev_id,
+                       "rx_worker_ring");
+       p->rx_worker_ring = qe_ring_create(buf, MAX_SW_PROD_Q_DEPTH,
+                       dev->data->socket_id);
+       if (p->rx_worker_ring == NULL) {
+               SW_LOG_ERR("Error creating RX worker ring for port %d\n",
+                               port_id);
+               return -1;
+       }
+
+       p->inflight_max = conf->new_event_threshold;
+
+       snprintf(buf, sizeof(buf), "sw%d_%s", dev->data->dev_id,
+                       "cq_worker_ring");
+       p->cq_worker_ring = qe_ring_create(buf, conf->dequeue_depth,
+                       dev->data->socket_id);
+       if (p->cq_worker_ring == NULL) {
+               qe_ring_destroy(p->rx_worker_ring);
+               SW_LOG_ERR("Error creating CQ worker ring for port %d\n",
+                               port_id);
+               return -1;
+       }
+       sw->cq_ring_space[port_id] = conf->dequeue_depth;
+
+       /* set hist list contents to empty */
+       for (i = 0; i < SW_PORT_HIST_LIST; i++) {
+               p->hist_list[i].fid = -1;
+               p->hist_list[i].qid = -1;
+       }
+       dev->data->ports[port_id] = p;
+
+       rte_smp_wmb();
+       p->initialized = 1;
+       return 0;
+}
+
+static void
+sw_port_release(void *port)
+{
+       struct sw_port *p = (void *)port;
+       if (p == NULL)
+               return;
+
+       qe_ring_destroy(p->rx_worker_ring);
+       qe_ring_destroy(p->cq_worker_ring);
+       memset(p, 0, sizeof(*p));
+}
+
 static int32_t
 qid_init(struct sw_evdev *sw, unsigned int idx, int type,
                const struct rte_event_queue_conf *queue_conf)
@@ -319,6 +398,8 @@ sw_probe(const char *name, const char *params)
                        .queue_setup = sw_queue_setup,
                        .queue_release = sw_queue_release,
                        .port_def_conf = sw_port_def_conf,
+                       .port_setup = sw_port_setup,
+                       .port_release = sw_port_release,
        };
 
        static const char *const args[] = {
index ddf0cd2..f5515e1 100644 (file)
 #define MAX_SW_PROD_Q_DEPTH 4096
 #define SW_FRAGMENTS_MAX 16
 
+/* report dequeue burst sizes in buckets */
+#define SW_DEQ_STAT_BUCKET_SHIFT 2
+/* how many packets pulled from port by sched */
+#define SCHED_DEQUEUE_BURST_SIZE 32
+
+#define SW_PORT_HIST_LIST (MAX_SW_PROD_Q_DEPTH) /* size of our history list */
+
 #define EVENTDEV_NAME_SW_PMD event_sw
 #define SW_PMD_NAME RTE_STR(event_sw)
 
@@ -129,12 +136,82 @@ struct sw_qid {
        uint8_t priority;
 };
 
+struct sw_hist_list_entry {
+       int32_t qid;
+       int32_t fid;
+       struct reorder_buffer_entry *rob_entry;
+};
+
+struct sw_evdev;
+
+struct sw_port {
+       /* new enqueue / dequeue API doesn't have an instance pointer, only the
+        * pointer to the port being enqueue/dequeued from
+        */
+       struct sw_evdev *sw;
+
+       /* set when the port is initialized */
+       uint8_t initialized;
+       /* A numeric ID for the port */
+       uint8_t id;
+
+       int16_t is_directed; /** Takes from a single directed QID */
+       /**
+        * For loadbalanced we can optimise pulling packets from
+        * producers if there is no reordering involved
+        */
+       int16_t num_ordered_qids;
+
+       /** Ring and buffer for pulling events from workers for scheduling */
+       struct qe_ring *rx_worker_ring __rte_cache_aligned;
+       /** Ring and buffer for pushing packets to workers after scheduling */
+       struct qe_ring *cq_worker_ring;
+
+       /* hole */
+
+       /* num releases yet to be completed on this port */
+       uint16_t outstanding_releases __rte_cache_aligned;
+       uint16_t inflight_max; /* app requested max inflights for this port */
+       uint16_t inflight_credits; /* num credits this port has right now */
+
+       uint16_t last_dequeue_burst_sz; /* how big the burst was */
+       uint64_t last_dequeue_ticks; /* used to track burst processing time */
+       uint64_t avg_pkt_ticks;      /* tracks average over NUM_SAMPLES burst */
+       uint64_t total_polls;        /* how many polls were counted in stats */
+       uint64_t zero_polls;         /* tracks polls returning nothing */
+       uint32_t poll_buckets[MAX_SW_CONS_Q_DEPTH >> SW_DEQ_STAT_BUCKET_SHIFT];
+               /* bucket values in 4s for shorter reporting */
+
+       /* History list structs, containing info on pkts egressed to worker */
+       uint16_t hist_head __rte_cache_aligned;
+       uint16_t hist_tail;
+       uint16_t inflights;
+       struct sw_hist_list_entry hist_list[SW_PORT_HIST_LIST];
+
+       /* track packets in and out of this port */
+       struct sw_point_stats stats;
+
+
+       uint32_t pp_buf_start;
+       uint32_t pp_buf_count;
+       uint16_t cq_buf_count;
+       struct rte_event pp_buf[SCHED_DEQUEUE_BURST_SIZE];
+       struct rte_event cq_buf[MAX_SW_CONS_Q_DEPTH];
+
+       uint8_t num_qids_mapped;
+};
+
 struct sw_evdev {
        struct rte_eventdev_data *data;
 
        uint32_t port_count;
        uint32_t qid_count;
 
+       /* Contains all ports - load balanced and directed */
+       struct sw_port ports[SW_PORTS_MAX] __rte_cache_aligned;
+
+       rte_atomic32_t inflights __rte_cache_aligned;
+
        /*
         * max events in this instance. Cached here for performance.
         * (also available in data->conf.nb_events_limit)
@@ -144,6 +221,9 @@ struct sw_evdev {
        /* Internal queues - one per logical queue */
        struct sw_qid qids[RTE_EVENT_MAX_QUEUES_PER_DEV] __rte_cache_aligned;
 
+       /* Cache how many packets are in each cq */
+       uint16_t cq_ring_space[SW_PORTS_MAX] __rte_cache_aligned;
+
        int32_t sched_quanta;
 
        uint32_t credit_update_quanta;